|Publication number||US6484123 B2|
|Application number||US 09/726,735|
|Publication date||Nov 19, 2002|
|Filing date||Nov 30, 2000|
|Priority date||Nov 30, 2000|
|Also published as||US20020099519|
|Publication number||09726735, 726735, US 6484123 B2, US 6484123B2, US-B2-6484123, US6484123 B2, US6484123B2|
|Inventors||Ashok N. Srivastava|
|Original Assignee||International Business Machines Corporation|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (2), Non-Patent Citations (3), Referenced by (60), Classifications (5), Legal Events (5)|
|External Links: USPTO, USPTO Assignment, Espacenet|
1. Field of the Invention
This invention generally relates to forecasting; and, more specifically, the invention relates to methods and systems to identify which predictors are important for making particular forecasts. The preferred embodiment of the invention may be used in any area where the collaborative filter (also called nearest neighbor) data mining technology is applied. Such areas include, but are not limited to, e-commerce, banking, manufacturing, securities trading, website personalization, mass marketing, communications, and medical diagnosis.
2. Prior art
The collaborative filter, also called the nearest neighbor model, is a mathematical model that is used to predict the value of a target variable given a set of input variables. For example, consider a scenario where a researcher has collected data regarding a population's age, height, weight, and gender. This information is contained in the following table:
Sample data containing a population's age,
height, weight, and gender information.
In this hypothetical example, suppose that the task is to take age, height, and weight information from a new table of data and use that information to determine the gender of the individual. Thus, suppose we are given a table as follows:
Sample data containing age, height, and weight
information for five new subjects. Gender is
The task is to apply the nearest neighbor algorithm to determine the gender of the subjects 6-10 based on the data in Tables 1 and 2.
In a simple implementation of the nearest neighbor algorithm, we compute the Euclidean distance between every subject in our test set (Table 2) and every subject in the training set (Table 1). The gender associated with the nearest match in Table 1 is assigned to the subjects in the test set. A variation on this algorithm is to use the average value of the K nearest neighbors. In this case, we take K=1. Thus:
Intersubject Euclidean distances. Cells
in bold correspond to the nearest neighbor.
The lowest value in each row indicates the nearest match between the subject in the test set and the subjects in the training set. Thus, Subject 6 is closest to Subject 4. The gender for Subject 6 is thus predicted to be male. Now, we can complete Table 2:
Subjects in Table 2 with Gender predictions.
A problem that arises in using this popular algorithm is that it is not known which variables led to the predictions made. For example, for subject 9, all we know at this point is that the nearest neighbor was subject 5. We do not know whether it was the age, height or weight of the subject that led to the conclusion that the subject's gender is female. At this point, the best we can say is that the ensemble of predictors (age, height and weight) led to this conclusion.
An object of this invention is to provide a novel method and system to identify which predictors are important for making a forecast.
Another object of the present invention is to determine which predictor, if any, are driving a particular predictor in a collaborative filter.
These and other objectives are attained with a method and system for identifying parameters that are important in predicting a target variable. The method comprises the steps of compiling training data, said training data identifying, for each of a first set of subjects, values for each of a first set of parameters; and compiling test data, said test data identifying, for each of a second set of subjects, values for each of a second set of parameters, said first and second sets of parameters having at least a plurality of common parameters.
The method comprises the further steps of using the data in the training data, and using a nearest neighbor procedure, to identify, for each of the second set of subjects, a value for a target parameter; and processing the training data and the test data, according to a predefined procedure, to determine the relative importance of at least selected ones of the first group of parameters in predicting the values for the target parameter.
Further benefits and advantages of the invention will become apparent from a consideration of the following detailed description, given with reference to the accompanying drawings, which specify and show preferred embodiments of the invention.
FIG. 1 generally outlines a preferred procedure embodying the invention.
FIGS. 2A and 2B describes more specifically a preferred procedure embodying the present invention.
FIG. 3 diagrammatically represents a procedure embodying the invention.
The present invention is a method and system to identify which predictors are important for making a forecast. Generally, in accordance with the invention, training data is compiled for a first set of subjects, test data is compiled for a second set of subjects, and the training data is used to identify target values for the second set of subjects. Also, the training data and the test data are processed to determine the relative importance of at least selected parameters in predicting the target values.
More specifically, with reference to FIG. 1, the training data identifies, for each of a first set of subjects, values for each of a first group of parameters. The test data identifies, for each of a second set of subjects, values for a second group of parameters. These first and second groups of parameters have at least a plurality of common parameters; however, the test data do not include values for at least one parameter that is in the first group of parameters. The training data is used, employing a nearest neighbor procedure, to identify, for each of the second set of subjects, a value for that one parameter for which the test data do not have values. This one parameter is referred to as the target parameter, and the values for this parameter are the target values.
With reference to FIGS. 2A, 2B and 3, the training data may be given in a table (or matrix) of data D of size (N×P), with each row corresponding to an observation and each column corresponding to a measurement. The columns of D correspond to the predictor variables. Also, a vector T of size (N×1) may be provided or used. Without the loss of generality, it may be noted, T may be a binary vector. The test data may be given in a matrix E of size (M×P), which contains measurements from the same sources as the matrix D.
A first problem is to predict the unknown targets U(M×1) vector associated with the matrix E. Procedures for solving this problem are known in the prior art, as discussed above. A second problem is, for each prediction, to determine which predictors are important in the prediction. The present invention provides a solution for this second problem.
At step 101, a value, j, is set equal to 1, which corresponds to the jth row of the matrix e. In step 102, the value K is set equal to the number of nearest neighbors used in the next neighbor algorithm to determine the target values. As will be understood by those of ordinary skill in the art, a wide range of values may be used for K. In order to improve the statistical significance of the procedure, K is preferably greater than about ten, and more preferably greater than 15. K may, however, have a value in the hundreds or even thousands.
At step 103, the nearest K neighbors, in matrix D, of the jth row of matrix E are identified. Any suitable nearest neighbor procedure, such as standard Euclidean measures, may be employed to do this. At step 104, a matrix Q (or Q(K,P)), is established. Q consists of K row or rows (depending on whether K is 1 or greater than 1, respectively). In particular, Q is the entire row or rows of the matrix D that is or are the nearest neighbor or neighbors of the jth row of matrix E.
At step 105, a matrix R (or R(a,b)) is calculated using the values from the E and Q matrices. Specifically, R is calculated according to the formula:
The matrix R will be of the size K×P.
At step 106, for each column in the matrix R, we compute the mean and standard deviations of the column. Then at step 107, for each column of R, we eliminate values of R that lie outside certain bounds. For instance, we may eliminate values of R that lie outside a W standard deviation bound from the mean, where W is chosen by the user. As an example, W may lie between two and six. Next, as represented by step 108, additional relatively large values may be eliminated from R. For instance, for each column of R, a mean of the remaining values of R may be calculated, and only those values kept that are below the mean—(W/2)* standard deviation. Next, the columns of R are ranked in decreasing order of the number (not the value) of the values remaining in each column.
Then the column of matrix D that corresponds to the column of R that has the greatest number of remaining values is identified. The correspondence between the columns of matrix R and the columns of matrix D is based on the column number. For example, the first column of matrix r corresponds to the first column of matrix D, the fifth column of matrix R corresponds to the fifth column of matrix D, and so forth. The parameter of the identified column of D is the most important predictor in forecasting the target value for the row j of matrix E.
Analogously, the parameter of the column of matrix D that corresponds to the column of matrix R that has the second largest number of remaining values, is the second most important predictor in forecasting the target value for the row J of matrix E, the parameter of the column of matrix D that corresponds to the column of matrix R that has the third largest number of remaining values, is the third most important predictor in forecasting the target value for the for j of matrix E, and so forth.
The above procedure may be repeated for each row j of matrix E, as represented by step 109.
The present invention has a wide variety of uses in various industries, including, for example, banking, finance, securities, and e-commerce. One very important use for this invention is in the e-commerce industry, since collaborative filters are extensively used for personalizing web-sites. An exemplary application of this technology to e-commerce follows.
For instance, a web-site may be constructed where consumer demographic data is gathered through a registration process, and consumer purchase behavior is recorded. In such a case, it may be desirable for the web-site to suggest automatically potential cross-sells to the customers. For example, if a consumer registers and divulges age, gender, income level, etc., and the consumer's purchases are recorded, then a collaborative filter could be applied to recommend that the consumer buy another product. The collaborative filter predicts which item is most likely to be of interest to the consumer based on demographics and purchase information.
In this example, without the technology disclosed herein, the collaborative filter would just recommend a product, but could not give a reason why the consumer should make the purchase. The present invention may be used to give reasons why the consumer should make the purchase. In particular, those reasons are the parameter or parameters identified, using the method and system disclosed herein, as the most important predictor or predictors in forecasting the product the consumer should purchase.
While it is apparent that the invention herein disclosed is well calculated to fulfill the objects stated above, it will be appreciated that numerous modifications and embodiments may be devised by those skilled in the art, and it is intended that the appended claims cover all such modifications and embodiments as fall within the true spirit and scope of the present invention.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US5485621 *||Apr 6, 1994||Jan 16, 1996||Siemens Corporate Research, Inc.||Interactive method of using a group similarity measure for providing a decision on which groups to combine|
|US6304675 *||Dec 28, 1993||Oct 16, 2001||Sandia Corporation||Visual cluster analysis and pattern recognition methods|
|1||"Data Mining The Data Mining Challenge: Turning Raw Data Into Business Gold", http://www.sas.com/software/data-mining, SAS Institute's Data Mining Solution, (last modified May 22, 2000).|
|2||"Enterprise Miner", http://www.sas.com/software/components/miner.htp, Enterprise Miner, (last modified May 22, 2000).|
|3||"IBM DB2 Intelligent Miner for Data", http://www.-4.ibm.com/software/data/iminer/fordata/index.htm, IBM Software: Database and Data Managem . . . B2 Intelligent Miner for Data: Overview, (last modified May 22, 2000).|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US7024381||May 14, 2003||Apr 4, 2006||Netflix, Inc.||Approach for renting items to customers|
|US7082407 *||Aug 19, 1999||Jul 25, 2006||Amazon.Com, Inc.||Purchase notification service for assisting users in selecting items from an electronic catalog|
|US7254552||Jan 30, 2004||Aug 7, 2007||Amazon.Com, Inc.||Notification service for assisting users in selecting items from an electronic catalog|
|US7403910||Jun 18, 2001||Jul 22, 2008||Netflix, Inc.||Approach for estimating user ratings of items|
|US7546252||Dec 23, 2003||Jun 9, 2009||Netflix, Inc.||Approach for managing rental items across a plurality of distribution locations|
|US7617127||Feb 6, 2004||Nov 10, 2009||Netflix, Inc.||Approach for estimating user ratings of items|
|US7676034||Mar 5, 2004||Mar 9, 2010||Wai Wu||Method and system for matching entities in an auction|
|US7685028||May 28, 2004||Mar 23, 2010||Gross John N||Method of testing inventory management/shipping systems|
|US7685117||Jun 7, 2004||Mar 23, 2010||Hayley Logistics Llc||Method for implementing search engine|
|US7689432||Jun 7, 2004||Mar 30, 2010||Hayley Logistics Llc||System and method for influencing recommender system & advertising based on programmed policies|
|US7742978||Apr 11, 2007||Jun 22, 2010||Swaptree, Inc.||Multi-transaction system and method|
|US7885849||Jun 7, 2004||Feb 8, 2011||Hayley Logistics Llc||System and method for predicting demand for items|
|US7890363||Jun 7, 2004||Feb 15, 2011||Hayley Logistics Llc||System and method of identifying trendsetters|
|US7894595||May 2, 2008||Feb 22, 2011||Wai Wu||Telephony control system with intelligent call routing|
|US7916858||Sep 18, 2006||Mar 29, 2011||Toby Heller||Agent training sensitive call routing system|
|US7966342||May 6, 2005||Jun 21, 2011||Hayley Logistics Llc||Method for monitoring link & content changes in web pages|
|US8065223||Jun 21, 2010||Nov 22, 2011||Swaptree, Inc.||Multi-transaction system and method|
|US8078615 *||Apr 11, 2003||Dec 13, 2011||Stumbleupon, Inc.||Method and system for single-action personalized recommendation and display of internet content|
|US8103540||Jun 7, 2004||Jan 24, 2012||Hayley Logistics Llc||System and method for influencing recommender system|
|US8140388||Jun 7, 2004||Mar 20, 2012||Hayley Logistics Llc||Method for implementing online advertising|
|US8180689||Nov 15, 2010||May 15, 2012||Amazon.Com, Inc.||Services for enabling users to share information regarding E-commerce transactions|
|US8249955||Mar 22, 2010||Aug 21, 2012||John Nicholas Gross||Method of testing item availability and delivery performance of an e-commerce site|
|US8300798||Apr 3, 2007||Oct 30, 2012||Wai Wu||Intelligent communication routing system and method|
|US8346673||May 10, 2004||Jan 1, 2013||Blockbuster L.L.C.||System and method for provisioning audiovisual works|
|US8433622||May 28, 2004||Apr 30, 2013||Media Queue, Llc||Method of controlling electronic commerce queue|
|US8606717||Jul 10, 2006||Dec 10, 2013||Media Queue, Llc||Playable media delivery capacity exchange method|
|US8612311||Jul 10, 2006||Dec 17, 2013||Media Queue, Llc||Hybrid distribution method for playable media|
|US8688462||Feb 2, 2004||Apr 1, 2014||Media Queue, Llc||Media auto exchange system and method|
|US8688503 *||Jun 22, 2012||Apr 1, 2014||American Express Travel Related Services Company, Inc.||System and method for targeting family members of transaction account product holders to receive supplementary transaction account products|
|US8700538||Feb 2, 2004||Apr 15, 2014||Media Queue, Llc||Media exchange system and method|
|US8712867||Feb 2, 2004||Apr 29, 2014||Media Queue, Llc||System for providing access to playable media|
|US8738541||Jun 22, 2004||May 27, 2014||Media Queue, Llc||Method of processing rental requests and returns|
|US8744929||Aug 20, 2007||Jun 3, 2014||Amazon.Com, Inc.||Services for enabling users to share information regarding E-commerce transactions|
|US8751307||Dec 1, 2011||Jun 10, 2014||Hayley Logistics Llc||Method for implementing online advertising|
|US8831205||Mar 15, 2013||Sep 9, 2014||Wai Wu||Intelligent communication routing|
|US8971519||Oct 30, 2013||Mar 3, 2015||Steven Hoffberg||Agent training sensitive call routing system|
|US20020052858 *||Dec 4, 2001||May 2, 2002||Insyst Ltd.||Method and tool for data mining in automatic decision making systems|
|US20030195884 *||Apr 11, 2003||Oct 16, 2003||Eric Boyd||Method and system for single-action personalized recommendation and display of internet content|
|US20040153413 *||Feb 2, 2004||Aug 5, 2004||Gross John N.||Notification system and method for media Queue|
|US20040158503 *||Feb 2, 2004||Aug 12, 2004||Gross John N.||Media queue monitor|
|US20040162783 *||Feb 2, 2004||Aug 19, 2004||Gross John N.||Media queue replenisher|
|US20040172274 *||Feb 2, 2004||Sep 2, 2004||Gross John N.||Media auto exchange system and method|
|US20040172275 *||Feb 2, 2004||Sep 2, 2004||Gross John N.||Media exchange system & method|
|US20040243479 *||May 28, 2004||Dec 2, 2004||Gross John N.||Method of monitoring electronic commerce queue|
|US20040243480 *||May 28, 2004||Dec 2, 2004||Gross John N.||Method of controlling electronic commerce queue|
|US20040243570 *||May 28, 2004||Dec 2, 2004||Gross John N.||Method of testing inventory management/shipping systems|
|US20040249700 *||Jun 7, 2004||Dec 9, 2004||Gross John N.||System & method of identifying trendsetters|
|US20040260574 *||Jun 7, 2004||Dec 23, 2004||Gross John N.||System and method for influencing recommender system & advertising based on programmed policies|
|US20040260600 *||Jun 7, 2004||Dec 23, 2004||Gross John N.||System & method for predicting demand for items|
|US20040267604 *||Jun 7, 2004||Dec 30, 2004||Gross John N.||System & method for influencing recommender system|
|US20040267669 *||Jun 22, 2004||Dec 30, 2004||Gross John N.||Method of processing rental requests and returns|
|US20050086127 *||Dec 23, 2003||Apr 21, 2005||Hastings W. R.||Approach for managing rental items across a plurality of distribution locations|
|US20050261987 *||Jan 30, 2004||Nov 24, 2005||Bezos Jeffrey P||Notification service for assisting users in selecting items from an electronic catalog|
|US20060015352 *||May 10, 2004||Jan 19, 2006||Wynn Dominic C||System and method for provisioning audiovisual works|
|US20060155575 *||Mar 6, 2006||Jul 13, 2006||Gross John N||Media delivery prioritization system and method|
|US20060212367 *||Jun 7, 2006||Sep 21, 2006||Gross John N||Method of selecting and distributing items to consumers of electronic media|
|US20060241967 *||Jul 10, 2006||Oct 26, 2006||Gross John N||Playable Media Delivery Capacity Exchange Method|
|US20060242269 *||Jul 10, 2006||Oct 26, 2006||Gross John N||Hybrid Distribution Method for Playable Media|
|US20070255624 *||Apr 14, 2006||Nov 1, 2007||Swaptree, Inc.||Automated Trading System and Method|
|US20080249843 *||Jun 16, 2008||Oct 9, 2008||Gross John N||Notification System and Method for Multiple Media Queues|
|U.S. Classification||702/181, 702/127|
|Nov 30, 2000||AS||Assignment|
|Jan 9, 2006||FPAY||Fee payment|
Year of fee payment: 4
|Jun 28, 2010||REMI||Maintenance fee reminder mailed|
|Nov 19, 2010||LAPS||Lapse for failure to pay maintenance fees|
|Jan 11, 2011||FP||Expired due to failure to pay maintenance fee|
Effective date: 20101119